Transformers and VIT

the main reason of building a transformer is for

• text summarization

• question answering

• translation

• masked word prediction

Bert train on a huge dataset to do ...

• masked word prediction

• next sentence prediction

• translation

• masked word prediction and nsp

a layer that encodes text into vectors

• positional encoding

• multi-head attention

• embedding layer

• residual connection

adds positions to words in a sentence

• positional encoding

• multi-head attention

• embedding layer

• residual connection

focus on some words using cosine similarity

• positional encoding

• multi-head attention

• embedding layer

• residual connection

depend on query, key, and value

• positional encoding

• multi-head attention

• embedding layer

• residual connection

to save the positional encoding from being gone we use ...

• positional encoding

• multi-head attention

• embedding layer

• residual connection

in decoder we hide future words using

• positional encoding

• masked multi-head attention

• embedding layer

• residual connection

in VIT we add cls to be used for

• classification

• predicting position 0 slice of the image

• predicting position 1 slice of the image

• next sentence prediction

to keep the order of the image parts correct we use

• positional encoding

• masked multi-head attention

• embedding layer

• residual connection